2. What’s a Table?
A dictionary is a set of (key, value) pairs.
• We’ll assume you can update the dictionary (insertions,
deletions, updates) and query the dictionary (point
queries, range queries)
• B-Trees and Fractal Trees are examples of dictionaries
• Hashes are not (range queries are not supported)
2
3. What’s a Table?
3
a b c
100 5 45
101 92 2
156 56 45
165 6 2
198 202 56
206 23 252
256 56 2
412 43 45
A table is a set of
dictionaries.
Example:
create table foo (a
int, b int, c int,
primary key(a));
Then we insert a bunch
of data and get...
4. What’s a Table?
The sort order of a dictionary is defined by the key
• For the data structures/storage engines we’ll think about,
range queries on the sort order are FAST
• Range queries on any other order require a table scan =
SLOW
• Point queries -- retrieving the value for one particular key --
is SLOW
‣A single point query is fast, but reading a bunch of rows this way is going
to be 2 orders of magnitude slower than reading the same number of
rows in range query order
4
5. What’s an Index?
A Index I on table T is itself a dictionary
• We need to define the (key, value) pairs.
• The key in index I is a subset of fields in the primary
dictionary T.
• The value in index I is the primary key in T.
‣ There are other ways to define the value, but we’re sticking with this.
Example:
alter table foo add key(b);
Then we get...
5
6. What’s an Index?
6
a b c
100 5 45
101 92 2
156 56 45
165 6 2
198 202 56
206 23 252
256 56 2
412 43 45
b a
5 100
6 165
23 206
43 412
56 156
56 256
92 101
202 198
Primary key(b)
7. Q: count(*) where a<120;
7
a b c
100 5 45
101 92 2
156 56 45
165 6 2
198 202 56
206 23 252
256 56 2
412 43 45
b a
5 100
6 165
23 206
43 412
56 156
56 256
92 101
202 198
8. Q: count(*) where a<120;
7
a b c
100 5 45
101 92 2
156 56 45
165 6 2
198 202 56
206 23 252
256 56 2
412 43 45
b a
5 100
6 165
23 206
43 412
56 156
56 256
92 101
202 198
100 5 45
101 92 2
2
9. Q: count(*) where b>50;
8
a b c
100 5 45
101 92 2
156 56 45
165 6 2
198 202 56
206 23 252
256 56 2
412 43 45
b a
5 100
6 165
23 206
43 412
56 156
56 256
92 101
202 198
10. Q: count(*) where b>50;
8
a b c
100 5 45
101 92 2
156 56 45
165 6 2
198 202 56
206 23 252
256 56 2
412 43 45
b a
5 100
6 165
23 206
43 412
56 156
56 256
92 101
202 198
56 156
56 256
92 101
202 198
4
11. Q: sum(c) where b>50;
9
a b c
100 5 45
101 92 2
156 56 45
165 6 2
198 202 56
206 23 252
256 56 2
412 43 45
b a
5 100
6 165
23 206
43 412
56 156
56 256
92 101
202 198
13. What are indexes good for?
Indexes make queries go fast
• Each index will speed up some subset of queries
Design indexes with queries in mind
• Pick most important queries and set up indexes for those
• Consider the cost of maintaining the indexes
10
14. What are indexes good for?
Indexes make queries go fast
• Each index will speed up some subset of queries
Design indexes with queries in mind
• Pick most important queries and set up indexes for those
• Consider the cost of maintaining the indexes
10
15. Goals of Talk
3 simple rules for designing good indexes
Avoid details of any data structure
• B-trees and Fractal Trees are interesting and fun for the
algorithmically minded computer scientist, but the 3 rules
will apply equally well to either data structure.
• All we need to care about is that range queries are fast
(per row) and that point queries are much slower (per
row).
11
16. The Rule to Rule Them All
There is no absolute rule
• Indexing is like a math problem
• Rules help, but each scenario is its own problem, which
requires problem-solving analysis
That said, rules help a lot
12
17. Three Basic Rules
1. Retrieve less data
• Less bandwidth, less processing, ...
2. Avoid point queries
• Not all data access cost is the same
• Sequential access is MUCH faster than random access
3. Avoid Sorting
• GROUP BY and ORDER BY queries do post-retrieval
work
• Indexing can help get rid of this work
13
19. Rule 1: Example of slow query
Example TABLE (1B rows, no indexes):
•create table foo (a int, b int, c int);
Query (1000 rows match):
•select sum(c) from foo where b=10 and a<150;
Query Plan:
• Rows where b=10 and a<150 can be anywhere in the table
• Without an index, entire table is scanned
Slow execution:
• Scan 1B rows just to count 1000 rows
20. Rule 1: How to add an index
What should we do?
• Reduce the data retrieved
• Analyze much less than 1B rows
How (for a simple select)?
• Design index by focusing on the WHERE clause
‣This defines what rows the query is interested in
‣Other rows are not important for this query
16
21. Rule 1: How to add an index
What should we do?
• Reduce the data retrieved
• Analyze much less than 1B rows
How (for a simple select)?
• Design index by focusing on the WHERE clause
‣This defines what rows the query is interested in
‣Other rows are not important for this query
16
select sum(c) from foo where b=10 and a<150;
22. Rule 1: Which index?
Option 1: key(a)
Option 2: key(b)
Which is better? Depends on selectivity:
• If there are fewer rows where a<150, then key(a) is better
• If there are fewer rows where b=10, then key(b) is better
Option 3: key(a) AND key(b), then MERGE
• We’ll come to this later
17
23. Rule 1: Picking the best key
Neither key(a) nor key(b) is optimal
Suppose:
• 200,000 rows exist where a<150
• 100,000 rows exist where b=10
• 1000 rows exist where b=10 and a<150
Then either index retrieves too much data
For better performance, indexes should try to
optimize over as many pieces of the where clause
as possible
• We need a composite index
18
24. Composite indexes reduce data retrieved
Where clause: b=5 and a<150
• Option 1: key(a,b)
• Option 2: key(b,a)
Which one is better?
• key(b,a)!
KEY RULE:
• When making a composite index, place equality checking
columns first. Condition on b is equality, but not on a.
19
25. Q: where b=5 and a>150;
20
a b c
100 5 45
101 6 2
156 5 45
165 6 2
198 6 56
206 5 252
256 5 2
412 6 45
b,a a
5,100 100
5,156 156
5,206 206
5,256 256
6,101 101
6,165 165
6,198 198
6,412 412
26. Q: where b=5 and a>150;
20
a b c
100 5 45
101 6 2
156 5 45
165 6 2
198 6 56
206 5 252
256 5 2
412 6 45
b,a a
5,100 100
5,156 156
5,206 206
5,256 256
6,101 101
6,165 165
6,198 198
6,412 412
5,156 156
5,206 206
5,256 256
27. Composite Indexes: No equality clause
What if where clause is:
•where a>100 and a<200 and b>100;
Which is better?
• key(a), key(b), key(a,b), key(b,a)?
KEY RULE:
• As soon as a column on a composite index is NOT used for
equality, the rest of the composite index no longer reduces
data retrieved.
‣key(a,b) is no better* than key(a)
‣key(b,a) is no better* than key(b)
21
28. Composite Indexes: No equality clause
What if where clause is:
•where a>100 and a<200 and b>100;
Which is better?
• key(a), key(b), key(a,b), key(b,a)?
KEY RULE:
• As soon as a column on a composite index is NOT used for
equality, the rest of the composite index no longer reduces
data retrieved.
‣key(a,b) is no better* than key(a)
‣key(b,a) is no better* than key(b)
21
* Are there corner cases where it helps? Yes, but rare.
29. Q: where b>=5 and a>150;
22
a b c
100 5 45
101 6 2
156 5 45
165 6 2
198 6 56
206 5 252
256 5 2
412 6 45
b,a a
5,100 100
5,156 156
5,206 206
5,256 256
6,101 101
6,165 165
6,198 198
6,412 412
32. Composite Indexes: Another example
WHERE clause: b=5 and c=100
• key(b,a,c) is as good as key(b), because a is not used
in clause, so having c in index doesn’t help. key(b,c,a)
would be much better.
23
a b c
100 5 100
101 6 200
156 5 200
165 6 100
198 6 100
206 5 200
256 5 100
412 6 100
b,a,c a
5,100,100 100
5,156,200 156
5,206,200 206
5,256,100 256
6,101,200 101
6,165,100 165
6,198,100 6,198
6,412,100 412
33. Composite Indexes: Another example
WHERE clause: b=5 and c=100
• key(b,a,c) is as good as key(b), because a is not used
in clause, so having c in index doesn’t help. key(b,c,a)
would be much better.
23
a b c
100 5 100
101 6 200
156 5 200
165 6 100
198 6 100
206 5 200
256 5 100
412 6 100
b,a,c a
5,100,100 100
5,156,200 156
5,206,200 206
5,256,100 256
6,101,200 101
6,165,100 165
6,198,100 6,198
6,412,100 412
5,100,100 100
5,156,200 156
5,206,200 206
5,256,100 256
34. Rule 1: Recap
Goal is to reduce rows retrieved
• Create composite indexes based on where clause
• Place equality-comparison columns at the beginning
• Make first non-equality column in index as selective as
possible
• Once first column in a composite index is not used for
equality, or not used in where clause, the rest of the
composite index does not help reduce rows retrieved
‣Does that mean they aren’t helpful?
‣They might be very helpful... on to Rule 2.
24
36. Rule 2: Avoid point queries
Table:
•create table foo (a int, b int, c int,
primary key(a), key(b));
Query:
•select sum(c) from foo where b>50;
Query plan: use key(b)
• retrieval cost of each row is high because random point
queries are done
26
37. Q: sum(c) where b>50;
27
a b c
100 5 45
101 92 2
156 56 45
165 6 2
198 202 56
206 23 252
256 56 2
412 43 45
b a
5 100
6 165
23 206
43 412
56 156
56 256
92 101
202 198
39. Rule 2: Avoid Point Queries
Table:
•create table foo (a int, b int, c int,
primary key(a), key(b));
Query:
•select sum(c) from foo where b>50;
Query plan: scan primary table
• retrieval cost of each row is CHEAP!
• But you retrieve too many rows
28
40. Q: sum(c) where b>50;
29
a b c
100 5 45
101 92 2
156 56 45
165 6 2
198 202 56
206 23 252
256 56 2
412 43 45
b a
5 100
6 165
23 206
43 412
56 156
56 256
92 101
202 198
42. Rule 2: Avoid Point Queries
Table:
•create table foo (a int, b int, c int,
primary key(a), key(b));
Query:
•select sum(c) from foo where b>50;
What if we add another index?
• What about key(b,c)?
• Since we index on b, we retrieve only the rows we need.
• Since the index has information about c, we don’t need to
go to the main table. No point queries!
30
45. Covering Index
An index covers a query if the index has enough
information to answer the query.
Examples:
Q: select sum(c) from foo where b<100;
Q: select sum(d) from foo where b<100;
Indexes:
• key(b,c) -- covering index for first query
• key(b,d) -- covering index for second query
• key(b,c,d) -- covering index for both
32
46. How to build a covering index
Add every field from the select
• Not just where clause
Q: select c,d from foo where a=10 and
b=100;
Mistake: add index(a,b);
• This doesn’t cover the query. You still need point queries
to retrieve c and d values.
Correct: add index(a,b,c,d);
• Includes all referenced fields
• Place a and b at beginning by Rule 1
33
47. What if Primary Key matches where?
Q: select sum(c) from foo where b>100
and b<200;
Schema: create table foo (a int, b
int, c int, ai int auto_increment,
primary key (b,ai));
• Query does a range query on primary dictionary
• Only one dictionary is accessed, in sequential order
• This is fast
Primary key covers all queries
• If sort order matches where clause, problems solved
34
48. What’s a Clustering Index
What if primary key doesn’t match the where
clause?
• Ideally, you should be able to declare secondary indexes
that carry all fields
• AFAIK, storage engines don’t let you do this
• With one exception... TokuDB
• TokuDB allows you to declare any index to be
CLUSTERING
• A CLUSTERING index covers all queries, just like the
primary key
35
49. Clustering Indexes in Action
Q: select sum(c) from foo where b<100;
Q: select sum(d) from foo where b>200;
Q: select c,e from foo where b=1000;
Indexes:
• key(b,c) covers first query
• key(b,d) covers second query
• key(b,c,e) covers first and third queries
• key(b,c,d,e) covers all three queries
Indexes require a lot of analysis of queries
36
50. Clustering Indexes in Action
Q: select sum(c) from foo where b<100;
Q: select sum(d) from foo where b>200;
Q: select c,e from foo where b=1000;
What covers all queries?
• clustering key(b)
Clustering keys let you focus on the where
clause
• They eliminate point queries and make queries fast
37
51. More on clustering: Index Merge
We had example:
• create table foo(a int, b int, c int);
• select sum(c) from foo where b=10 and
a<150;
Suppose
• 200,000 rows have a<150
• 100,000 rows have b=10
• 1000 rows have b=10 and a<150
What if we use key(a) and key(b) and merge
results?
38
52. Query Plans
Merge plan:
• Scan 200,000 rows in key(a) where a<150
• Scan 100,000 rows in key(b) where b=10
• Merge the results and find 1000 row identifiers that
match query
• Do point queries with 1000 row identifiers to retrieve c
Better than no indexes
• Reduces number of rows scanned compared to no index
• Reduces number of point queries compared to not
merging
39
53. Does Clustering Help Merging?
Suppose key(a) is clustering
Query plan:
• Scan key(a) for 200,000 rows where a<150
• Scan resulting rows for ones where b=10
• Retrieve c values from 1000 remaining rows
Once again, no point queries
What’s even better?
• clustering key(b,a)!
40
54. Rule 2 Recap
Avoid Point Queries
Make sure index covers query
• By mentioning all fields in the query, not just those in the
where clause
Use clustering indexes
• Clustering indexes cover all queries
• Allows user to focus on where clause
• Speeds up more (and unforeseen) queries -- simplifies
database design
41
56. Rule 3: Avoid Sorting
Simple selects require no post-processing
• select * from foo where b=100;
Just get the data and return to user
More complex queries do post-processing
• GROUP BY and ORDER BY sort the data
Index selection can avoid this sorting step
43
57. Avoid Sorting
Q1: select count(c) from foo;
Q2: select count(c) from foo group by
b, order by b;
Q1 plan:
• While doing a table scan, count rows with c
Q2 plan
• Scan table and write data to a temporary file
• Sort tmp file data by b
• Rescan sorted data, counting rows with c, for each b
44
58. Avoid Sorting
Q2: select count(c) from foo group by
b, order by b;
Q2: what if we use key(b,c)?
• By adding all needed fields, we cover query. FAST!
• By sorting first by b, we avoid sort. FAST!
Take home:
• Sort index on group by or order by fields to avoid
sorting
45
63. Example
select count(*) from foo where c=5,
group by b;
key(c,b):
• First have c to limit rows retrieved (R1)
• Then have remaining rows sorted by b to avoid sort (R3)
• Remaining rows will be sorted by b because of equality
test on c
48
66. Example
select sum(d) from foo where c=100,
group by b;
key(c,b,d):
• First have c to limit rows retrieved (R1)
• Then have remaining rows sorted by b to avoid sort (R3)
• Make sure index covers query, to avoid point queries (R2)
49
67. Example
Sometimes, there is no clear answer
• Best index is data dependent
Q: select count(*) from foo where
c<100, group by b;
Indexes:
• key(c,b)
• key(b,c)
50
68. Example
Q: select count(*) from foo where
c<100, group by b;
Query plan for key(c,b):
• Will filter rows where c<100
• Still need to sort by b
‣ Rows retrieved will not be sorted by b
‣ where clause does not do an equality check on c, so b values are scattered in clumps for
different c values
51
69. Example
Q: select count(*) from foo where
c<100, group by b;
Query plan for key(b,c):
• Sorted by b, so R3 is covered
• Rows where c>=100 are also processed, so not taking
advantage of R1
52
70. Example
Which is better?
• Answer depends on the data
• If there are many rows where c>=100, saving time by not
retrieving these useless rows helps. Use key(c,b).
• If there aren’t so many rows where c>=100, the time to
execute the query is dominated by sorting. Use
key(b,c).
The point is, in general, rules of thumb help, but
often they help us think about queries and
indexes, rather than giving a recipe.
53
71. Why not just load up
with indexes?
Need to keep up with insertion load
More indexes = smaller max load
72. Indexing cost
Space
• Issue
‣Each index adds storage requirements
• Options
‣Use compression (i.e., aggressive compression of 5-15x always on for
TokuDB)
Performance
• Issue
‣B-trees while fast for certain indexing tasks (in memory, sequential keys),
are over 20x slower for other types of indexing
• Options
‣Fractal Tree indexes (TokuDB’s data structure) is fast at all kinds of
indexing (i.e., random keys, large tables, wide keys, etc...)
‣ No need to worry about what type of index you are creating.
‣ Fractal trees enable customers to index early, index often
55
73. Last Caveat
Range Query Performance
• Issue
‣Rule #2 (range query performance over point query performance)
depends on range queries being fast
‣However B-trees can get fragmented
‣ from deletions, from random insertions, ...
‣ Fragmented B-trees get slow for range queries
• Options
‣For B-trees, optimize tables, dump and reload, (ie, time consuming and
offline maintenance) ...
‣For Fractal Tree indexing (TokuDB), not an issue
‣ Fractal trees don’t fragment
56
74. Thanks!
For more information...
• Please contact me at zardosht@tokutek.com for any
thoughts or feedback
• Please visit Tokutek.com for a copy of this presentation
(goo.gl/S2LBe) to learn more about the power of
indexing, read about Fractal Tree indexes, or to download
a free eval copy of TokuDB
57